The buzz about ChatGPT
Will ChatGPT eliminate technical writing as we know it?
Social Media is abuzz with the capabilities of ChatGPT released by OpenAI last week. In fact, the usage of ChatGPT became so popular that OpenAI had to increase the capacity allocated to this service. Praise for ChatGPT has ranged from users mentioning that ChatGPT could replace Google Search to people wondering how could academic essays ever be graded anymore after the advent of services like ChatGPT. Many people have compared academic articles written by them many years ago to the output from ChatGPT and have found the output from ChatGPT superior.
So what’s ChatGPT?
Natural Language Models such as GPT-3 which we had written about in a previous post write sentences by filling in the best possible tokens based on an originating paragraph of text. You can think of them as engines that fill in the blanks with the most appropriate set of words. However, GPT-3 has been shown to have two main limitations: i) It lacks memory, and so you cannot ask it to give alternate explanations to questions, and ii) it lacks a feedback loop, therefore lacking the ability to distinguish poor text such as those that make racial and hate comments, or incite violence from good text.
ChatGPT makes Natural Language Models conversational. It improves on GPT-3, or rather GPT-3.5 (the latest GPT-3 model from OpenAI) in the following ways. First, the training of GPT-3.5 is improved by using input from human labelers who choose the best series of tokens from the outputs of the original GPT-3.5 model. Then a Reinforcement Learning (RL) model is used to train a policy with a reward model based on the choices made by a human labeller who rank the various outputs they had in conversations with ChatGPT. The policy is continuously updated to let ChatGPT improve on the quality of results with time.
The result is that ChatGPT performs much better in answering questions and also refusing to answer questions about hate or violence. The OpenAI page on ChatGPT gives a good summary about many of the details. You should try to create an OpenAI account and try it yourself.
On our end, we asked ChatGPT about the Vector robot. Please see the answers below, and comment on how you find the quality of answers from ChatGPT.